US8068373B1 - Power management of memory via wake/sleep cycles - Google Patents

Power management of memory via wake/sleep cycles Download PDF

Info

Publication number
US8068373B1
US8068373B1 US12/911,181 US91118110A US8068373B1 US 8068373 B1 US8068373 B1 US 8068373B1 US 91118110 A US91118110 A US 91118110A US 8068373 B1 US8068373 B1 US 8068373B1
Authority
US
United States
Prior art keywords
memory
power
dimm
dma
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
US12/911,181
Inventor
George Totolos, Jr.
Scott M. Westbrook
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
NetApp Inc
Original Assignee
Network Appliance Inc
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Network Appliance Inc filed Critical Network Appliance Inc
Priority to US12/911,181 priority Critical patent/US8068373B1/en
Application granted granted Critical
Publication of US8068373B1 publication Critical patent/US8068373B1/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F1/00Details not covered by groups G06F3/00 - G06F13/00 and G06F21/00
    • G06F1/26Power supply means, e.g. regulation thereof
    • G06F1/32Means for saving power
    • G06F1/3203Power management, i.e. event-based initiation of a power-saving mode
    • G06F1/3234Power saving characterised by the action undertaken
    • G06F1/325Power saving in peripheral device
    • G06F1/3275Power saving in memory, e.g. RAM, cache
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F13/00Interconnection of, or transfer of information or other signals between, memories, input/output devices or central processing units
    • G06F13/14Handling requests for interconnection or transfer
    • G06F13/20Handling requests for interconnection or transfer for access to input/output bus
    • G06F13/28Handling requests for interconnection or transfer for access to input/output bus using burst mode transfer, e.g. direct memory access DMA, cycle steal
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02DCLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00Energy efficient computing, e.g. low power processors, power management or thermal management
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02DCLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00Reducing energy consumption in communication networks
    • Y02D30/50Reducing energy consumption in communication networks in wire-line communication networks, e.g. low power modes or reduced link rate

Definitions

  • At least one embodiment of the present invention pertains to a method of managing computer memory to reduce its power consumption, and more particularly, to reducing the power consumption of a computer memory while minimizing adverse performance effects.
  • Random access storage contrasts with sequential access storage media, where access to one portion of the stored information may take significantly longer than access to another portion. For example, if information is stored on a magnetic tape, then retrieving information at the end of the tape may take much longer than retrieving information at the beginning.
  • Random access capability is often critical to adequate performance of applications on a general- or special-purpose computer.
  • memory access patterns may be perfectly predictable, or may be sufficiently regular that even a significant time penalty to begin accessing data can be tolerated by amortizing that penalty over a large number of operations.
  • the processor will never need to access only a single byte within a block of memory, but will only need to move the entire block from one place to another. In these and other similar situations, memory need not provide random access capabilities.
  • DIMM dual in-line memory module
  • DIMMs require electrical power to perform their storage and retrieval functions.
  • the amount of power consumed during operation is generally proportional to factors such as the amount of storage provided and the speed of operation. For example, one 2 GB DIMM available on the market requires a current of 1.71 A at 1.8V (3.08 W) when storing or retrieving data.
  • the memory consumes 2.26 W.
  • Other memory modules may consume different amounts of power.
  • DRAM Standard dynamic RAM
  • SDRAM synchronous DRAM
  • DDR double data rate
  • DDR1 double data rate version 2
  • DDR2 double data rate version 2
  • self-refresh is typically used by portable devices such as laptops when they are put into a “sleep” state for the sake of battery conservation.
  • self-refresh mode the aforementioned 2 GB DIMM requires just 180 ma (324 mW), a power savings of almost 90%.
  • Other memory modules may achieve greater or lesser power savings in self-refresh mode.
  • a dynamic memory in self-refresh mode is different from a non-volatile memory such as EEPROM or Flash that can retain its contents even when power is completely removed.
  • a DRAM in self-refresh mode continues to consume a small amount of power, and will lose its contents if power is interrupted.
  • the power savings offered by self-refresh mode come at a price. New information cannot be written to, and existing data cannot be retrieved from, a DIMM in self-refresh mode. If these functions of the memory are required, the “sleeping” memory must be “awoken:” it must go through a power-up procedure that may take many clock cycles, or that may restrict the memory to a lower-speed operational mode in lieu of those clock cycles.
  • the time-consuming power-up procedure is necessary to permit the memory to synchronize its internal operations with an external clock by means of a delay-locked loop (“DLL”).
  • DLL delay-locked loop
  • the synchronization may take significantly longer than the normal memory access time of only a few clock cycles, possibly exceeding the access time by a factor of 10 or more.
  • some memories can be powered up and ready for operation within a few cycles (without performing synchronization), but in that event, access cycles may be limited to only a fraction of the full speed possible if the synchronization is performed.
  • DIMM power management operations have referred specifically to DIMMs that conform to the JEDEC DDR1 and DDR2 standards, it is expected that future versions of these standards (e.g. DDR3, DDR4, and so forth) will maintain similar power management capabilities and time or performance penalties for using them. Thus, the subsequent discussions will remain applicable to future DIMMs that implement descendants of DDR1 and DDR2.
  • other (non-JEDEC) memory modules may offer a “sleep” mode wherein the memory retains its contents, but cannot perform any of its other functions until a “wake” procedure is performed. The sleep mode will typically be associated with reduced power consumption, and the wake procedure will often impose a time or performance penalty.
  • the subsequent discussions are relevant to this sort of memory module as well.
  • Embodiments of the invention manage power to a plurality of memory modules, where each module has at least a low- and a high-power state.
  • a set of memory access requests is maintained, each request specifying a transfer of data to or from one of the memory modules.
  • a request from the set is selected, and the memory module that is involved in the request is changed from low-power state to high-power state. Then the request is executed, and the memory module is returned to its low-power state.
  • FIG. 1 shows an environment including a server that can benefit from embodiments of the invention.
  • FIG. 2 is a block diagram of the server of FIG. 1 , showing the components that may be found in such a server.
  • FIG. 3 is a flow chart showing the steps and decisions involved in an embodiment of the invention.
  • FIG. 4 is a timeline showing how proper coordination of events occurring within a memory bank can reduce the impact of long delays in memory access.
  • FIG. 5 shows the functional blocks that may be contained in a hardware embodiment of the invention.
  • Embodiments of the invention manage the power supplied to the DIMMs that comprise a computer memory.
  • the memory is idle (i.e., when the system is not currently engaged in transferring data to or from the memory)
  • all DIMMs that are managed are placed in a low power (or power saving) state such as “self-refresh” or “sleep” mode.
  • a request to transfer data to or from a DIMM in the memory occurs, that DIMM is powered up (“woken”) so that the request can be serviced. After the transfer is finished, the DIMM is returned to the low power state.
  • a memory may be considered “large” by reference to other parameters of the system of which the memory is a part. For example, a memory that would normally consume more power than the system can provide might be considered “large.”
  • Embodiments of the invention may be useful in any system where there are multiple DIMMs (or other types of memory modules) in use, and reducing the amount of power used by those DIMMs provides an overall system benefit.
  • the expressions “low power” and “high power” as used herein are different from the terms “low” and “high” that are sometimes used in describing the states of a binary circuit.
  • the expressions refer to the overall amount of electrical power consumed by a memory module. This amount of power is generally only weakly related to the data stored in the module. In other words, a memory containing all zeros (or logical “low” states) generally will not consume a significantly different amount of power than a similar memory containing all ones (or logical “high” states).
  • FIG. 1 shows a network environment containing computers 110 and 120 , which act as clients of server 140 .
  • the clients communicate with the server through a medium such as network 130 .
  • Computer system 160 may be useful to configure, monitor, or manage server 140 .
  • the server may control a number of mass storage devices such as disks 150 .
  • Server 140 may provide data storage services to its clients.
  • server 140 may have extensive storage resources, protected by data redundancy, backup, and power reliability measures.
  • Data storage services may be provided on a per-file basis, where server 140 appears as to its clients as a network file server, or on a per-block basis, where server 140 appears to the clients as a directly-attached disk drive. In either situation, server 140 may need to move large quantities of data about, both internally and between its clients and its mass storage devices.
  • server 140 may not need to examine or modify individual bits or bytes, but may simply be required to accept or deliver blocks consisting of, for example, 1,024 (1K) bytes, 4,096 (4K) bytes, or other numbers of bytes of user data.
  • This storage server application may be implemented by a computer system such as that depicted in functional-block form in FIG. 2 .
  • Central processing units (CPUs) 210 may control and coordinate the operation of various devices connected to system bus 240 , under the direction of instructions stored in memory 220 .
  • Disk 230 may contain other instructions and data that are not immediately required by the system in its operations.
  • the system can communicate with its clients through network adapters 260 , and with its mass storage devices such as disks 150 through storage adapters 270 .
  • the system may also benefit from additional cache memory, separate from memory 220 , that may be installed on an add-in card such as PCI cache memory card 250 .
  • the storage server may place user data on mass storage devices 150 for long-term availability, but to improve performance and throughput, it may be beneficial to store, or “cache,” copies of the user data in memory 220 .
  • memory 220 may become full of these cached copies of user data, and a least-recently-used (“LRU”) or similar algorithm may be employed to determine which cached data to remove when it is desired to store copies of new data. Blocks of data selected for removal are called “victims,” and would normally be discarded from memory 220 to make room for the new data. (The discarded victim data can, of course, be read in again from mass storage devices 150 if necessary.
  • LRU least-recently-used
  • PCI cache memory 250 a secondary cache such as PCI cache memory 250 .
  • the secondary cache permits more user data to be stored for ready access—data in PCI cache memory 250 may be located and copied into memory 220 for further processing by CPUs 210 much faster than the same data could be located and transferred from mass storage devices 150 .
  • DMA direct memory access
  • a CPU may configure a special-purpose circuit with source and destination addresses and other information, and then the circuit automatically generates addresses and performs consecutive memory cycles to move data from one place to another. Note that the memory access patterns resulting from a DMA transfer are predictable given knowledge of the information with which the special-purpose circuit is configured. Furthermore, in the specific example of a storage server being considered here, many DMA transfers will move blocks of data consisting of, for example, 4K bytes.
  • the memory will include a number of DIMMs (or equivalent memory devices in other packages, where each device has a power saving state) that are initially in a low-power mode.
  • circuitry such as a memory controller associated with the memory may perform operations according to the flowchart of FIG. 3 , which describes one embodiment of the invention.
  • the flowchart shows three processes that can be performed mostly concurrently by hardware, firmware, or software entities operating largely independently.
  • the memory controller is idle, all processes are inactive, no data are being transferred between the memory bank and the host system, and all of the DIMMs may be in a sleep state.
  • DMA descriptor prefetch is concerned with receiving or obtaining DMA requests from the host system and providing the requests to the other processes.
  • a DMA request arrives ( 315 )
  • it is stored or queued ( 320 )
  • memory control process 330 may be notified that there is work to do ( 325 ).
  • memory control process 330 selects a DMA descriptor ( 335 ) and identifies the DIMM to (or from) which data are to be transferred ( 340 ). If the identified DIMM is not currently active ( 345 ), the memory control process 330 will start the power-up procedure ( 350 ) and there may be a delay ( 355 ) while the DIMM becomes ready. Once the DIMM has reached its normal or operational power state and is ready to transfer data, the DMA request can be issued to the DMA engine 370 for execution ( 360 ).
  • DMA engine 370 can accept the request ( 375 ) and begin transferring data ( 380 ). Data are moved between the DIMM and another device or component on the system in portions which may be four, eight, sixteen, or another number of bytes in size, as supported by the hardware, as shown by the looping notation at block 380 .
  • the DMA engine may check for another request that requires access to the same DIMM ( 385 ). If there is such a request, it can be accepted and started immediately. Otherwise, the DMA engine may return the just-accessed DIMM to “sleep” mode ( 390 ) and resume waiting for a DMA request on a DIMM that is ready.
  • FIG. 4 shows the temporal relationship between the steps and decisions described with reference to FIG. 3 , and illustrates how most memory “wake” delays can be overlapped with other processing to avoid impacting overall system performance.
  • DIMM 1 becomes operational and the first DMA transfer occurs during time period 430 .
  • the circuitry selects another DMA request that may have been scheduled, determines which DIMM contains the source or destination memory address, and begins the power-up sequence of that DIMM (“DIMM 2”).
  • the process of selecting a DMA request, powering up the affected DIMM, performing the DMA transfer, and powering down the DIMM can be repeated indefinitely during time period 450 .
  • the last scheduled DMA operation may be completed (time period 460 ), and the memory module returns to idle mode during time period 470 .
  • the process described permits the memory to be operated at only a fraction of the full power required by all the DIMMs in normal operating state because at any time, one DIMM is operating, one DIMM is transitioning to its normal power state, and one DIMM is (rapidly) switching from normal power state to low-power state. The remaining DIMMs are in low-power state. Furthermore, in this example, each DMA transfer takes longer than the power-up time of the next DIMM, so the power-up latency time can be completely hidden and does not affect the overall system latency or response time. The only exception is the first power-up cycle, time period 320 , when the memory bank begins operating again after an idle period.
  • FIG. 5 shows a block diagram of the structural components that may be present in a computer memory system such as a portion of main memory 220 in FIG. 2 or an add-in memory card such as PCI cache memory 250 in FIG. 2 , which includes an embodiment of the invention.
  • the cache control functions represented within dashed line 500 may be implemented by one or more integrated circuits, programmable gate arrays, or other circuit elements.
  • Host bus 501 represents a communication channel that permits the cache controller to transfer data between DIMM banks 590 and other components in the system (not shown in this figure).
  • Peripheral Component Interface—Extended (“PCI-X”) Interface Core 504 , PCI-X Manager 508 and Command-Status Register 512 provide electrical and timing control appropriate for interfacing to host bus 501 .
  • the DMA Descriptor Prefetch section 516 receives DMA requests from the system and provides them to DMA engine 532 .
  • the DMA engine 532 including DMA control 536 and byte aligner 540 , moves data between the host system and memory control block 544 .
  • the source (or destination) on the host system may be, for example, memory 220 as shown in FIG. 2 , or some other device or component that can be accessed through host bus 501 .
  • Memory control block including built-in self test (“BIST”) engine 548 , memory controller 552 , scrubber 556 , command and data FIFOs 560 , 564 and 568 , error correction code (“ECC”) generator 572 and ECC checker 576 , and DDR2 cores 580 and 584 , moves data in or out of DIMMs 590 , and performs the power coordination functions of embodiments of the invention.
  • memory controller 552 may provide appropriate signals to bring a DIMM that is the source or target of a DMA operation out of low-power mode, and may provide other signals to place a DIMM that has completed a DMA operation back into low-power mode.
  • the cache controller 500 shown in FIG. 5 is a programmable memory controller that accesses memory based on DMA descriptors it receives from the host system. It manages the power supplied to DIMMs 590 according to an embodiment of the invention, so that the DIMMs use less power than they would if each DIMM was in fully operational status at all times. This management is essential in a system where the overall power consumption is limited to a maximum value. For example, the maximum allowed power consumption of a PCI card is 25 W.
  • a PCI cache memory card could normally hold no more than 20 GB (9 DIMMs in “ready” mode at 2.26 W and 1 DIMM operating at 3.08 W, with 1.58 W excess for miscellaneous control functions).
  • the DIMMs on the PCI card were operated according to an embodiment of the invention, then by a similar calculation the card could contain 116 GB: 56 DIMMs in self-refresh mode at 324 mW, 1 DIMM powering up at 2.26 W and 1 DIMM operating at 3.08 W, with 1.52 W for miscellaneous and control functions.
  • the time to activate a DIMM from low-power mode is assumed to be 200 clock cycles.
  • the time to complete a DMA transfer is assumed to be 256 cycles.
  • the time required to perform calculations such as determining which DIMM a DMA transfer targets, and the time required to issue power-down and power-up commands, is less than 56 cycles.
  • the power-up latency depends upon the specific memory module in use, and may be greater or less than 200 clock cycles. In subsequent paragraphs, the latency will simply be called T power-up . Likewise, the time to switch a memory from operational mode to low-power mode will be called T power-down .
  • DMA transfers require 256 cycles comes from the premise that such transfers will move 4K blocks over a 64-bit-wide bus, moving 16 bytes (128 bits) each clock cycle. Naturally, transfers of smaller blocks or transfers over a wider bus can be completed in less time.
  • DMA operations are a clear case where memory access patterns are predictable, other applications might exhibit similarly coherent locality of reference. For example, a video display driver might require access to known blocks of display memory, or a 3-D rendering engine might operate on vertex or shading data stored in contiguous memory locations. The time required to perform the DMA operation or to access the memory in a DIMM for other purposes will be called T access .
  • Administrative tasks also include such operations as selecting a next memory access request to perform from among the set of access requests issued.
  • the access requests may be stored by a scheme as simple as a first-in, first-out (FIFO) queue.
  • FIFO first-in, first-out
  • Embodiments of the present invention can hide power-management delays when performing a sequence of work as long as T access >T admin +T power-up T power-down .
  • an embodiment of the invention may sort access requests according to an algorithm such as the known elevator algorithm, so that several access operations can be performed on one DIMM while the next DIMM is powering up.
  • a request would be selected not based upon an order in which it was received, but based upon the DIMM to or from which data were to be transferred. Thus, multiple requests to access the same DIMM would be processed together.
  • embodiments of the invention may permit the use of more memory than would otherwise be possible under a set of power constraints, so that the overall application benefit exceeds the cost of unhidden delays.
  • Some embodiments of the invention may perform special processing to select the first DMA operation to perform after an idle period, to take advantage of a DIMM's ability to operate in an unsynchronized mode. As discussed earlier, if a block of memory can be transferred at low speed faster than it can be transferred at high speed (after taking into account T power-up ), then it may be preferable to perform the first operation at low speed. An embodiment of the invention may select such a DMA operation as the first one to perform after an idle period. After this first operation, the memory controller can continue as described above.

Abstract

A method of managing power states of memory modules while performing memory access operations is disclosed. Memory modules are in a power saving state until an access operation involving the module is to be performed. The module is placed in an operational mode, then the access operation is performed, then the module is returned to the power saving state. Apparatus and systems using the method are also disclosed and claimed.

Description

CLAIM OF PRIORITY
This application claims priority to co-pending U.S. patent application Ser. No. 11,691,321 by George Totolos and Scott M. Westbrook, filed on 26 Mar. 2007 and allowed on 10 Jun. 2010. This application also claims priority to U.S. patent application Ser. No. 11/118,505 by George Totolos and Scott M. Westbrook, filed 28 Apr. 2005 and allowed on 23 Feb. 2007 as U.S. Pat. No. 7,218,566.
FIELD OF THE INVENTION
At least one embodiment of the present invention pertains to a method of managing computer memory to reduce its power consumption, and more particularly, to reducing the power consumption of a computer memory while minimizing adverse performance effects.
BACKGROUND
Computer processors and other devices (often operating under the control of a processor) frequently need to store or retrieve information in a computer memory. Memories usually permit access to any portion of the information they contain within an essentially constant amount of time. This property is indicated by the words “random access” in the name commonly applied to a type of computer memory, “random access memory” (“RAM”). Other types of memory also permit random access, though their names may indicate other important properties of the memory. For example, read-only memory (“ROM”), including electrically-erasable, programmable read-only memory (“EEPROM” or “Flash”) can also be accessed randomly. Random access storage contrasts with sequential access storage media, where access to one portion of the stored information may take significantly longer than access to another portion. For example, if information is stored on a magnetic tape, then retrieving information at the end of the tape may take much longer than retrieving information at the beginning.
Random access capability is often critical to adequate performance of applications on a general- or special-purpose computer. However, in some circumstances, memory access patterns may be perfectly predictable, or may be sufficiently regular that even a significant time penalty to begin accessing data can be tolerated by amortizing that penalty over a large number of operations. Also, in some applications, it may be known that the processor will never need to access only a single byte within a block of memory, but will only need to move the entire block from one place to another. In these and other similar situations, memory need not provide random access capabilities.
Whether computer memory is accessed randomly or sequentially, and as individual bytes or as larger blocks, its storage and retrieval functions are typically performed by electronic integrated circuits. The circuits are often packaged in a form known as a dual in-line memory module, or “DIMM.” This term will be used herein, but it should be recognized that the discussion applies equally to any memory device having the characteristics described, and not only to memories in the specific DIMM package.
DIMMs require electrical power to perform their storage and retrieval functions. The amount of power consumed during operation is generally proportional to factors such as the amount of storage provided and the speed of operation. For example, one 2 GB DIMM available on the market requires a current of 1.71 A at 1.8V (3.08 W) when storing or retrieving data. When not operating, but capable of responding immediately to a read or write command, the memory consumes 2.26 W. Other memory modules may consume different amounts of power.
Standard dynamic RAM (“DRAM”) devices available today (including synchronous DRAM, “SDRAM;” double data rate, “DDR” or “DDR1;” and double data rate version 2, “DDR2” devices) generally have a low power mode defined by the JEDEC Solid State Technology Association (formerly Joint Electron Device Engineering Council). This mode is called “self-refresh,” and is typically used by portable devices such as laptops when they are put into a “sleep” state for the sake of battery conservation. In self-refresh mode, the aforementioned 2 GB DIMM requires just 180 ma (324 mW), a power savings of almost 90%. Other memory modules may achieve greater or lesser power savings in self-refresh mode. A dynamic memory in self-refresh mode is different from a non-volatile memory such as EEPROM or Flash that can retain its contents even when power is completely removed. A DRAM in self-refresh mode continues to consume a small amount of power, and will lose its contents if power is interrupted.
Of course, the power savings offered by self-refresh mode come at a price. New information cannot be written to, and existing data cannot be retrieved from, a DIMM in self-refresh mode. If these functions of the memory are required, the “sleeping” memory must be “awoken:” it must go through a power-up procedure that may take many clock cycles, or that may restrict the memory to a lower-speed operational mode in lieu of those clock cycles. The time-consuming power-up procedure is necessary to permit the memory to synchronize its internal operations with an external clock by means of a delay-locked loop (“DLL”). In some systems, the synchronization may take significantly longer than the normal memory access time of only a few clock cycles, possibly exceeding the access time by a factor of 10 or more. Alternatively, some memories can be powered up and ready for operation within a few cycles (without performing synchronization), but in that event, access cycles may be limited to only a fraction of the full speed possible if the synchronization is performed.
In either case, switching DIMMs into and out of self-refresh mode to save power is not a promising strategy when the system's memory access patterns are not predictable, because the power-up time or performance penalties may be incurred whenever a DIMM containing required information is in low-power mode.
Although these discussions of DIMM power management operations have referred specifically to DIMMs that conform to the JEDEC DDR1 and DDR2 standards, it is expected that future versions of these standards (e.g. DDR3, DDR4, and so forth) will maintain similar power management capabilities and time or performance penalties for using them. Thus, the subsequent discussions will remain applicable to future DIMMs that implement descendants of DDR1 and DDR2. Also, other (non-JEDEC) memory modules may offer a “sleep” mode wherein the memory retains its contents, but cannot perform any of its other functions until a “wake” procedure is performed. The sleep mode will typically be associated with reduced power consumption, and the wake procedure will often impose a time or performance penalty. The subsequent discussions are relevant to this sort of memory module as well.
SUMMARY OF THE INVENTION
Embodiments of the invention manage power to a plurality of memory modules, where each module has at least a low- and a high-power state. A set of memory access requests is maintained, each request specifying a transfer of data to or from one of the memory modules. A request from the set is selected, and the memory module that is involved in the request is changed from low-power state to high-power state. Then the request is executed, and the memory module is returned to its low-power state.
BRIEF DESCRIPTION OF DRAWINGS
Embodiments of the invention are illustrated by way of example and not by way of limitation in the accompanying figures. In those figures, like reference numbers indicate similar elements.
FIG. 1 shows an environment including a server that can benefit from embodiments of the invention.
FIG. 2 is a block diagram of the server of FIG. 1, showing the components that may be found in such a server.
FIG. 3 is a flow chart showing the steps and decisions involved in an embodiment of the invention.
FIG. 4 is a timeline showing how proper coordination of events occurring within a memory bank can reduce the impact of long delays in memory access.
FIG. 5 shows the functional blocks that may be contained in a hardware embodiment of the invention.
DETAILED DESCRIPTION
Consider an application in which a large amount of memory is required to fulfill a product need. If the application's memory access patterns are predictable, rather than random, then it may be possible to operate the large amount of memory according to an embodiment of the invention and either dramatically reduce the power consumed by the memory, or dramatically increase the amount of memory that can be operated within a fixed power budget. A specific application with appropriate memory access patterns will be described first, then the concepts will be generalized to assist in identifying other applications that can benefit from the invention. Finally, extensions and alternate embodiments will be discussed.
Embodiments of the invention manage the power supplied to the DIMMs that comprise a computer memory. When the memory is idle (i.e., when the system is not currently engaged in transferring data to or from the memory), all DIMMs that are managed are placed in a low power (or power saving) state such as “self-refresh” or “sleep” mode. When a request to transfer data to or from a DIMM in the memory occurs, that DIMM is powered up (“woken”) so that the request can be serviced. After the transfer is finished, the DIMM is returned to the low power state.
Of course, no single limit can be identified, beyond which a memory is “large,” and less than which it is “small.” Instead, a memory may be considered “large” by reference to other parameters of the system of which the memory is a part. For example, a memory that would normally consume more power than the system can provide might be considered “large.” Embodiments of the invention may be useful in any system where there are multiple DIMMs (or other types of memory modules) in use, and reducing the amount of power used by those DIMMs provides an overall system benefit.
Note that the expressions “low power” and “high power” as used herein are different from the terms “low” and “high” that are sometimes used in describing the states of a binary circuit. Here, the expressions refer to the overall amount of electrical power consumed by a memory module. This amount of power is generally only weakly related to the data stored in the module. In other words, a memory containing all zeros (or logical “low” states) generally will not consume a significantly different amount of power than a similar memory containing all ones (or logical “high” states).
FIG. 1 shows a network environment containing computers 110 and 120, which act as clients of server 140. The clients communicate with the server through a medium such as network 130. Computer system 160 may be useful to configure, monitor, or manage server 140. The server may control a number of mass storage devices such as disks 150.
Server 140 may provide data storage services to its clients. For example, server 140 may have extensive storage resources, protected by data redundancy, backup, and power reliability measures. Data storage services may be provided on a per-file basis, where server 140 appears as to its clients as a network file server, or on a per-block basis, where server 140 appears to the clients as a directly-attached disk drive. In either situation, server 140 may need to move large quantities of data about, both internally and between its clients and its mass storage devices. Furthermore, server 140 may not need to examine or modify individual bits or bytes, but may simply be required to accept or deliver blocks consisting of, for example, 1,024 (1K) bytes, 4,096 (4K) bytes, or other numbers of bytes of user data.
This storage server application may be implemented by a computer system such as that depicted in functional-block form in FIG. 2. Central processing units (CPUs) 210 may control and coordinate the operation of various devices connected to system bus 240, under the direction of instructions stored in memory 220. Disk 230 may contain other instructions and data that are not immediately required by the system in its operations. The system can communicate with its clients through network adapters 260, and with its mass storage devices such as disks 150 through storage adapters 270. The system may also benefit from additional cache memory, separate from memory 220, that may be installed on an add-in card such as PCI cache memory card 250.
In operation, the storage server may place user data on mass storage devices 150 for long-term availability, but to improve performance and throughput, it may be beneficial to store, or “cache,” copies of the user data in memory 220. Eventually, memory 220 may become full of these cached copies of user data, and a least-recently-used (“LRU”) or similar algorithm may be employed to determine which cached data to remove when it is desired to store copies of new data. Blocks of data selected for removal are called “victims,” and would normally be discarded from memory 220 to make room for the new data. (The discarded victim data can, of course, be read in again from mass storage devices 150 if necessary. The memory caching described here only affects system performance, not data integrity.)
Improved system performance may be possible if, instead of simply discarding cache victim data blocks, they are copied to a secondary cache such as PCI cache memory 250. The secondary cache permits more user data to be stored for ready access—data in PCI cache memory 250 may be located and copied into memory 220 for further processing by CPUs 210 much faster than the same data could be located and transferred from mass storage devices 150.
To transfer data to and from PCI cache memory 250, direct memory access (“DMA”) operations may be most efficient. In a DMA transfer, a CPU may configure a special-purpose circuit with source and destination addresses and other information, and then the circuit automatically generates addresses and performs consecutive memory cycles to move data from one place to another. Note that the memory access patterns resulting from a DMA transfer are predictable given knowledge of the information with which the special-purpose circuit is configured. Furthermore, in the specific example of a storage server being considered here, many DMA transfers will move blocks of data consisting of, for example, 4K bytes.
If the source or destination of a DMA transfer is a bank of memory operated according to an embodiment of the invention, the memory will include a number of DIMMs (or equivalent memory devices in other packages, where each device has a power saving state) that are initially in a low-power mode. When one or more DMA transfers are required, circuitry such as a memory controller associated with the memory may perform operations according to the flowchart of FIG. 3, which describes one embodiment of the invention.
The flowchart shows three processes that can be performed mostly concurrently by hardware, firmware, or software entities operating largely independently. When the memory controller is idle, all processes are inactive, no data are being transferred between the memory bank and the host system, and all of the DIMMs may be in a sleep state.
Process 310, DMA descriptor prefetch, is concerned with receiving or obtaining DMA requests from the host system and providing the requests to the other processes. When a DMA request arrives (315), it is stored or queued (320), and memory control process 330 may be notified that there is work to do (325).
When the DMA descriptor prefetch 310 has one or more DMA requests to service, memory control process 330 selects a DMA descriptor (335) and identifies the DIMM to (or from) which data are to be transferred (340). If the identified DIMM is not currently active (345), the memory control process 330 will start the power-up procedure (350) and there may be a delay (355) while the DIMM becomes ready. Once the DIMM has reached its normal or operational power state and is ready to transfer data, the DMA request can be issued to the DMA engine 370 for execution (360).
When a DMA request is pending and the source or destination DIMM is ready, DMA engine 370 can accept the request (375) and begin transferring data (380). Data are moved between the DIMM and another device or component on the system in portions which may be four, eight, sixteen, or another number of bytes in size, as supported by the hardware, as shown by the looping notation at block 380. When all the data have been transferred, the DMA engine may check for another request that requires access to the same DIMM (385). If there is such a request, it can be accepted and started immediately. Otherwise, the DMA engine may return the just-accessed DIMM to “sleep” mode (390) and resume waiting for a DMA request on a DIMM that is ready.
FIG. 4 shows the temporal relationship between the steps and decisions described with reference to FIG. 3, and illustrates how most memory “wake” delays can be overlapped with other processing to avoid impacting overall system performance.
In period 410, before t=0, the memory is idle and all the DIMMs are in a low-power state. At t=0, a first DMA request is scheduled. The circuitry determines which DIMM contains the source or destination memory addresses (the “target addresses”), and begins the power-up sequence of that DIMM. Other DMA requests may be scheduled during time period 420 (or, indeed, at any other time shown in the timeline). At t=200, DIMM 1 becomes operational and the first DMA transfer occurs during time period 430. In addition, the circuitry selects another DMA request that may have been scheduled, determines which DIMM contains the source or destination memory address, and begins the power-up sequence of that DIMM (“DIMM 2”). If the first DMA operation takes, for example, 256 clock cycles, then at t=456, the first DMA will complete. Since the power-up procedure of DIMM 2 takes only 200 cycles, the second DMA operation can be started immediately after the first DMA operation completes. The second DMA occurs during time period 440, and the circuitry may select another DMA request, determine which DIMM contains the source or destination memory address, and begin the power-up sequence of that DIMM (“DIMM 3”). Note that after a DMA transfer has moved data to or from a DIMM, that DIMM can be returned to low-power mode. The process of shutting down a DIMM takes only a few clock cycles, and is indicated by gray triangles 480 in the timeline.
The process of selecting a DMA request, powering up the affected DIMM, performing the DMA transfer, and powering down the DIMM can be repeated indefinitely during time period 450. Eventually, the last scheduled DMA operation may be completed (time period 460), and the memory module returns to idle mode during time period 470.
The process described permits the memory to be operated at only a fraction of the full power required by all the DIMMs in normal operating state because at any time, one DIMM is operating, one DIMM is transitioning to its normal power state, and one DIMM is (rapidly) switching from normal power state to low-power state. The remaining DIMMs are in low-power state. Furthermore, in this example, each DMA transfer takes longer than the power-up time of the next DIMM, so the power-up latency time can be completely hidden and does not affect the overall system latency or response time. The only exception is the first power-up cycle, time period 320, when the memory bank begins operating again after an idle period. However, in a busy system, such idle periods should be rare; conversely, if a system is so underutilized that idle periods are common, then any delay a client perceives due to a DIMM's power-up cycle may be shorter than the delay it might perceive due to a heavier load on the storage server.
Furthermore, as discussed previously, it may be possible to operate a DIMM that has just been brought out of low-power mode within only a few clock cycles (bypassing the time-consuming DLL synchronization process) by using slower, unsynchronized memory cycles. If, for example, the unsynchronized memory cycle takes twice as long as a synchronized cycle, then DMA transfers that would take fewer than 200 cycles at full speed can be completed faster by performing unsynchronized, half-speed accesses (requiring less than 400 cycles) than by spending 200 cycles synchronizing the DIMM clocks and performing the accesses at full speed. Some memory access patterns may make this option more attractive than suffering the power-up delay penalty at the beginning of a sequence of memory operations after an idle period. Note that even if the first access after an idle period is performed at the slower, unsynchronized speed, the second and subsequent accesses can be performed at full speed because the synchronization periods of those target DIMMs can be overlapped with the first memory operation or the previous memory operation.
FIG. 5 shows a block diagram of the structural components that may be present in a computer memory system such as a portion of main memory 220 in FIG. 2 or an add-in memory card such as PCI cache memory 250 in FIG. 2, which includes an embodiment of the invention. The cache control functions represented within dashed line 500 may be implemented by one or more integrated circuits, programmable gate arrays, or other circuit elements. Host bus 501 represents a communication channel that permits the cache controller to transfer data between DIMM banks 590 and other components in the system (not shown in this figure). Peripheral Component Interface—Extended (“PCI-X”) Interface Core 504, PCI-X Manager 508 and Command-Status Register 512 provide electrical and timing control appropriate for interfacing to host bus 501. The DMA Descriptor Prefetch section 516, including descriptor fetch controller 520, low-priority descriptor first-in, first-out (“FIFO”) 524 and high-priority descriptor FIFO 528, receives DMA requests from the system and provides them to DMA engine 532. The DMA engine 532, including DMA control 536 and byte aligner 540, moves data between the host system and memory control block 544. The source (or destination) on the host system may be, for example, memory 220 as shown in FIG. 2, or some other device or component that can be accessed through host bus 501.
Memory control block, including built-in self test (“BIST”) engine 548, memory controller 552, scrubber 556, command and data FIFOs 560, 564 and 568, error correction code (“ECC”) generator 572 and ECC checker 576, and DDR2 cores 580 and 584, moves data in or out of DIMMs 590, and performs the power coordination functions of embodiments of the invention. For example, memory controller 552 may provide appropriate signals to bring a DIMM that is the source or target of a DMA operation out of low-power mode, and may provide other signals to place a DIMM that has completed a DMA operation back into low-power mode.
The cache controller 500 shown in FIG. 5 is a programmable memory controller that accesses memory based on DMA descriptors it receives from the host system. It manages the power supplied to DIMMs 590 according to an embodiment of the invention, so that the DIMMs use less power than they would if each DIMM was in fully operational status at all times. This management is essential in a system where the overall power consumption is limited to a maximum value. For example, the maximum allowed power consumption of a PCI card is 25 W. Using the example power consumption for the 2 G DIMM mentioned above, one might calculate that a PCI cache memory card could normally hold no more than 20 GB (9 DIMMs in “ready” mode at 2.26 W and 1 DIMM operating at 3.08 W, with 1.58 W excess for miscellaneous control functions). However, if the DIMMs on the PCI card were operated according to an embodiment of the invention, then by a similar calculation the card could contain 116 GB: 56 DIMMs in self-refresh mode at 324 mW, 1 DIMM powering up at 2.26 W and 1 DIMM operating at 3.08 W, with 1.52 W for miscellaneous and control functions.
In the application described above, several assumptions are made. First, the time to activate a DIMM from low-power mode is assumed to be 200 clock cycles. Second, the time to complete a DMA transfer is assumed to be 256 cycles. Third, although it is not specifically discussed, it is assumed that the time required to perform calculations such as determining which DIMM a DMA transfer targets, and the time required to issue power-down and power-up commands, is less than 56 cycles. These assumptions will be examined below.
The power-up latency depends upon the specific memory module in use, and may be greater or less than 200 clock cycles. In subsequent paragraphs, the latency will simply be called Tpower-up. Likewise, the time to switch a memory from operational mode to low-power mode will be called Tpower-down.
The assumption that DMA transfers require 256 cycles comes from the premise that such transfers will move 4K blocks over a 64-bit-wide bus, moving 16 bytes (128 bits) each clock cycle. Naturally, transfers of smaller blocks or transfers over a wider bus can be completed in less time. Although DMA operations are a clear case where memory access patterns are predictable, other applications might exhibit similarly coherent locality of reference. For example, a video display driver might require access to known blocks of display memory, or a 3-D rendering engine might operate on vertex or shading data stored in contiguous memory locations. The time required to perform the DMA operation or to access the memory in a DIMM for other purposes will be called Taccess.
The “administrative” overhead required to identify the DIMM among a number of DIMMs that contains the source or target of a DMA transfer can reasonably be assumed to be small, as it typically involves only examining a subset of the bits of the DMA source and destination addresses. Administrative tasks also include such operations as selecting a next memory access request to perform from among the set of access requests issued. The access requests may be stored by a scheme as simple as a first-in, first-out (FIFO) queue. However, even complex queuing schemes can usually be performed within only a few tens of cycles, and in any case these calculations can be performed while the next DIMM is powering up. Administrative time will be called Tadmin.
Embodiments of the present invention can hide power-management delays when performing a sequence of work as long as Taccess>Tadmin+Tpower-up Tpower-down. In addition, even if a single access operation does not last long enough to completely overlap the administrative, power-up, and power-down times, an embodiment of the invention may sort access requests according to an algorithm such as the known elevator algorithm, so that several access operations can be performed on one DIMM while the next DIMM is powering up. In this embodiment, a request would be selected not based upon an order in which it was received, but based upon the DIMM to or from which data were to be transferred. Thus, multiple requests to access the same DIMM would be processed together. Finally, even if it is not always possible to hide all power-management delays, embodiments of the invention may permit the use of more memory than would otherwise be possible under a set of power constraints, so that the overall application benefit exceeds the cost of unhidden delays.
Some embodiments of the invention may perform special processing to select the first DMA operation to perform after an idle period, to take advantage of a DIMM's ability to operate in an unsynchronized mode. As discussed earlier, if a block of memory can be transferred at low speed faster than it can be transferred at high speed (after taking into account Tpower-up), then it may be preferable to perform the first operation at low speed. An embodiment of the invention may select such a DMA operation as the first one to perform after an idle period. After this first operation, the memory controller can continue as described above.
Embodiments of the present invention have been described largely by reference to specific examples and in terms of particular allocations of functionality to certain hardware and/or software components. However, those of skill in the art will recognize that the methods can also be performed by software and hardware that distribute the functions of embodiments of this invention differently than herein described. Such variations and implementations are understood to be apprehended according to the following claims. It should be noted that references to “an” or “one” embodiment in this disclosure are not necessarily to the same embodiment, and such references mean “at least one.”

Claims (5)

1. A system comprising:
a first memory;
a second memory; and
a programmable memory controller to transfer data between the first memory and the second memory according to a descriptor issued to the programmable memory controller, each memory having a sleep state and an operative state,
wherein the second memory is in the sleep state before the descriptor is issued to the programmable memory controller and after data are transferred between the first memory and the second memory, and
wherein, at all times,
no more than one memory is in the operative state,
no more than one memory is transitioning from the sleep state to the operative state, and
no more than one memory is transitioning from the operative state to the sleep state.
2. The system of claim 1, wherein the second memory in the sleep state consumes more than zero watts (0 W) and less power than the second memory in the operative state.
3. The system of claim 1, wherein the second memory maintains its contents in the sleep state.
4. The system of claim 1, wherein the second memory requires a first amount of time to change from the sleep state to the operative state, and a second amount of time to access a byte in the second memory, the first amount of time exceeding the second amount of time.
5. The system of claim 4, wherein the first amount of time exceeds the second amount of time by at least a factor of 10.
US12/911,181 2010-10-25 2010-10-25 Power management of memory via wake/sleep cycles Active US8068373B1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US12/911,181 US8068373B1 (en) 2010-10-25 2010-10-25 Power management of memory via wake/sleep cycles

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
US12/911,181 US8068373B1 (en) 2010-10-25 2010-10-25 Power management of memory via wake/sleep cycles

Publications (1)

Publication Number Publication Date
US8068373B1 true US8068373B1 (en) 2011-11-29

Family

ID=44994364

Family Applications (1)

Application Number Title Priority Date Filing Date
US12/911,181 Active US8068373B1 (en) 2010-10-25 2010-10-25 Power management of memory via wake/sleep cycles

Country Status (1)

Country Link
US (1) US8068373B1 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9298602B2 (en) 2013-11-27 2016-03-29 Lenovo (Singapore) Pte. Ltd. Nonvolatile random access memory use
US20160093400A1 (en) * 2014-09-30 2016-03-31 Apple Inc. Power-aware memory self-test unit

Citations (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6065124A (en) 1997-03-15 2000-05-16 Samsung Electronics Co., Ltd. Computer system having power saving and management function and method of controlling the same
US6118719A (en) 1998-05-20 2000-09-12 International Business Machines Corporation Self-initiated self-refresh mode for memory modules
US20020083288A1 (en) 2000-10-25 2002-06-27 Inode Technology, Inc. Addressing system for use in storage devices
US6571333B1 (en) 1999-11-05 2003-05-27 Intel Corporation Initializing a memory controller by executing software in second memory to wakeup a system
US6618791B1 (en) 2000-09-29 2003-09-09 Intel Corporation System and method for controlling power states of a memory device via detection of a chip select signal
US20040078524A1 (en) 2002-10-16 2004-04-22 Robinson John T. Reconfigurable cache controller for nonuniform memory access computer systems
US6781911B2 (en) 2002-04-09 2004-08-24 Intel Corporation Early power-down digital memory device and method
US20040260957A1 (en) 2003-06-20 2004-12-23 Jeddeloh Joseph M. System and method for selective memory module power management
US20050125703A1 (en) 2003-12-03 2005-06-09 International Business Machines Corporation Method and system for power management including local bounding of device group power consumption
US20050125702A1 (en) 2003-12-03 2005-06-09 International Business Machines Corporation Method and system for power management including device controller-based device use evaluation and power-state control
US20060020765A1 (en) 2004-07-02 2006-01-26 Peter Mahrla Configuration of components for a transition from a low-power operating mode to a normal-power operating mode
US20060164904A1 (en) 2005-01-24 2006-07-27 Hugo Saleh Dynamic pre-charge level control in semiconductor devices
US7721011B1 (en) * 2005-05-09 2010-05-18 Oracle America, Inc. Method and apparatus for reordering memory accesses to reduce power consumption in computer systems

Patent Citations (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6065124A (en) 1997-03-15 2000-05-16 Samsung Electronics Co., Ltd. Computer system having power saving and management function and method of controlling the same
US6118719A (en) 1998-05-20 2000-09-12 International Business Machines Corporation Self-initiated self-refresh mode for memory modules
US6571333B1 (en) 1999-11-05 2003-05-27 Intel Corporation Initializing a memory controller by executing software in second memory to wakeup a system
US6618791B1 (en) 2000-09-29 2003-09-09 Intel Corporation System and method for controlling power states of a memory device via detection of a chip select signal
US7069409B2 (en) 2000-10-25 2006-06-27 Grape Technology, Inc. System for addressing a data storage unit used in a computer
US20020083288A1 (en) 2000-10-25 2002-06-27 Inode Technology, Inc. Addressing system for use in storage devices
US6781911B2 (en) 2002-04-09 2004-08-24 Intel Corporation Early power-down digital memory device and method
US20040078524A1 (en) 2002-10-16 2004-04-22 Robinson John T. Reconfigurable cache controller for nonuniform memory access computer systems
US6961821B2 (en) 2002-10-16 2005-11-01 International Business Machines Corporation Reconfigurable cache controller for nonuniform memory access computer systems
US20040260957A1 (en) 2003-06-20 2004-12-23 Jeddeloh Joseph M. System and method for selective memory module power management
US20050125703A1 (en) 2003-12-03 2005-06-09 International Business Machines Corporation Method and system for power management including local bounding of device group power consumption
US20050125702A1 (en) 2003-12-03 2005-06-09 International Business Machines Corporation Method and system for power management including device controller-based device use evaluation and power-state control
US7155623B2 (en) 2003-12-03 2006-12-26 International Business Machines Corporation Method and system for power management including local bounding of device group power consumption
US20060020765A1 (en) 2004-07-02 2006-01-26 Peter Mahrla Configuration of components for a transition from a low-power operating mode to a normal-power operating mode
US20060164904A1 (en) 2005-01-24 2006-07-27 Hugo Saleh Dynamic pre-charge level control in semiconductor devices
US7721011B1 (en) * 2005-05-09 2010-05-18 Oracle America, Inc. Method and apparatus for reordering memory accesses to reduce power consumption in computer systems

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9298602B2 (en) 2013-11-27 2016-03-29 Lenovo (Singapore) Pte. Ltd. Nonvolatile random access memory use
US20160093400A1 (en) * 2014-09-30 2016-03-31 Apple Inc. Power-aware memory self-test unit
US9589672B2 (en) * 2014-09-30 2017-03-07 Apple Inc. Power-aware memory self-test unit

Similar Documents

Publication Publication Date Title
US7218566B1 (en) Power management of memory via wake/sleep cycles
CN109155143B (en) Fine granularity refresh
KR102519019B1 (en) Ordering of memory requests based on access efficiency
US8788777B2 (en) Memory on-demand, managing power in memory
CN107408079B (en) Memory controller with coherent unit for multi-level system memory
US20120030396A1 (en) Decoupled Memory Modules: Building High-Bandwidth Memory Systems from Low-Speed Dynamic Random Access Memory Devices
US11500797B2 (en) Computer memory expansion device and method of operation
JP6276470B2 (en) System and method for reducing standby power of volatile memory in portable computing devices
CN105518784A (en) Data processor with memory controller for high reliability operation and method
US11100013B2 (en) Scheduling of read and write memory access requests
US11682445B2 (en) Memory context restore, reduction of boot time of a system on a chip by reducing double data rate memory training
US11561862B2 (en) Refresh management for DRAM
CN111684427A (en) Cache control aware memory controller
US11748034B2 (en) Signalling for heterogeneous memory systems
EP3142015A1 (en) Low-power memory-access method and associated apparatus
WO2005069148A2 (en) Memory management method and related system
US7650472B2 (en) Method for reducing memory power consumption
US8484418B2 (en) Methods and apparatuses for idle-prioritized memory ranks
CN115516563A (en) Refresh management for DRAM
US8068373B1 (en) Power management of memory via wake/sleep cycles
US20200117265A1 (en) Speculative exit from power down mode of a dynamic random access memory rank
US10884477B2 (en) Coordinating accesses of shared resources by clients in a computing device
CN117136351A (en) Data texture clock switching
KR20230160854A (en) Memory controller power states
CN115729860A (en) Processing system, processing method and electronic equipment

Legal Events

Date Code Title Description
STCF Information on status: patent grant

Free format text: PATENTED CASE

FPAY Fee payment

Year of fee payment: 4

MAFP Maintenance fee payment

Free format text: PAYMENT OF MAINTENANCE FEE, 8TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1552); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

Year of fee payment: 8

MAFP Maintenance fee payment

Free format text: PAYMENT OF MAINTENANCE FEE, 12TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1553); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

Year of fee payment: 12